LLaDA-Image Launches Open Image Generation Stack

The compelling part is not simply another image model, but an unusually transparent attempt to expose how a competitive generator is trained. Its open recipe could matter more to researchers than its benchmark score, provided the promised training code actually arrives.

• –A 6B DiT paired with a frozen LLaDA2.0-Mini-based vision-language module unifies generation and reference-image editing in one model family.
• –Image-only pretraining before language alignment offers a practical alternative to relying entirely on massive image-text datasets.
• –LLaDA-Image-Turbo’s 2–4-step inference makes local experimentation and interactive editing substantially more viable.
• –The reported 53.53 English and 53.38 Chinese Qwen-Image-Bench scores are strong, but independent reproduction matters before treating the claimed open-source lead as settled.
• –Apache-2.0 checkpoints, FP8 variants, bilingual text rendering, and Diffusers support give developers a credible path to self-hosted deployment. [Hugging Face](https://huggingface.co/inclusionAI/LLaDA-Image-Turbo)